Errors and ambiguity at every level, so certainty has to flow both ways. In clean problems, a pipeline is fine: each stage resolves its layer and hands up a fact. In speech, no stage can resolve its own layer alone. The acoustics genuinely cannot decide between “seven” and “Devon” — but the grammar knows a digit is likely here, and that top-down knowledge is exactly what the bottom-up stage needed. A pipeline structurally forbids that conversation: by the time the parser exists, the phoneme decision is frozen. Low-confidence guesses compound stage over stage, and the final output is built on the rubble. (The other options are real 1970s engineering pains, but none of them is architectural — a faster, bigger machine fixes them and still can’t send grammatical knowledge downward.)
The blackboard’s answer is bidirectional inference through a shared hypothesis space: a strong word hypothesis raises the plausibility of the phonemes that would compose it, exactly as strong phonetics raises candidate words. Hearsay-II’s signature tactic was the island of certainty: find the stretch where some level is confident — anywhere in the utterance, any level of abstraction — and grow outward from it, letting each secured hypothesis constrain its noisy neighbors. Understanding spreads from strongholds instead of marching left-to-right. (Notice the human echo, deliberately: you do this when parsing a bad phone connection — catch “…meeting…” and “…Thursday…” and reconstruct the sentence from its islands. Blackboard theory began as a model of perception.)
What the design buys, beyond speech:
- Radical modularity. KSs share only the board’s hypothesis format. Add a new expert — Hearsay-II gained and shed KSs throughout its life — without touching the others. It’s lesson 3’s rule-modularity, promoted from rules to subsystems.
- Anytime behavior. The board always holds a current best interpretation with confidences; stop early and you get a usable partial answer, not a stack trace.
- Graceful degradation. A weak KS costs quality, not collapse — others route around it.
The cost, honestly: the intelligence doesn’t disappear, it moves into control — deciding which KS acts next, when hundreds could. Lesson 3’s moral (arbitration is behavior) returns at full scale, and it’s the next lesson’s entire subject.
For your harness: the correspondence is direct — shared scratchpad/context = blackboard; tools and sub-agents = KSs; orchestrator = control. The two ideas most worth stealing are islands of certainty (let confidence, not sequence, drive the plan — start where the task is most tractable and let results constrain the rest) and hypotheses-with-support-links instead of bare facts on the board, so conclusions can be re-rated when their support erodes. When a fixed tool-chain keeps building confidently on step 2’s shaky output, you are re-living 1974; that is pipeline rubble, and this lesson is the fix.
Where this goes: inside the machine — hypothesis levels, the scheduler and its agenda, and how BB1 put control knowledge itself on a blackboard.